Sample deliverable — differential expression, done two ways

The single most consequential decision in a multi-subject single-cell study is how differential expression is tested. This is that decision, run both ways on the same cells, on public data (GEO GSE96583 (Kang et al. 2018) — 8 donors, control vs IFN-β) — so you can see exactly how much the answer moves.

1 · The design

Eight donors, each sampled with and without interferon-β. The real sample size is eight per group, not 2,658 cells per group — and the two methods differ entirely in which of those numbers they believe.

Donor	Condition	Cells aggregated
101	ctrl	202
101	stim	244
1015	ctrl	779
1015	stim	653
1016	ctrl	369
1016	stim	320
1039	ctrl	126
1039	stim	161
107	ctrl	209
107	stim	164
1244	ctrl	415
1244	stim	284
1256	ctrl	390
1256	stim	370
1488	ctrl	295
1488	stim	336

2 · Both methods, same cells

Left: counts aggregated per donor and condition, then DESeq2 with donor in the model (a paired design). Right: every cell treated as an independent observation, Wilcoxon, BH-corrected — the default in most tutorials.

Comparison	Genes
Called by both methods	781
Only by the per-cell test	345
Only by pseudobulk	1,336
Disagreement	68.3% of all calls

The usual warning about per-cell tests is that they inflate the gene list. Here — a very strong, very consistent treatment effect with a paired design — the honest result is more interesting than the cliché: the per-cell test returns fewer genes overall, uniquely calls 345 that the design-aware analysis does not support, and misses 1,336 that it does. Two thirds of the final list depends on a methodological choice that is rarely stated in a paper.

3 · Pseudobulk result

4 · The correctness check

Interferon-β has a known signature. All ten canonical interferon-stimulated genes come out significantly up in the pseudobulk analysis — 10/10 — with fold changes in the range you would expect. A method that cannot recover a known biology should not be trusted with an unknown one, and this is the check we build into every project.

Gene	Pseudobulk log2FC	Pseudobulk padj	Per-cell log2FC
ISG15	7.74	2.1e-86	8.76
IFI6	4.98	6.8e-35	5.23
IFIT1	8.64	1.7e-45	8.8
IFIT3	7.85	5.2e-88	8.25
MX1	6.23	1.2e-53	6.27
OAS1	5.22	2.7e-52	5.26
IFI44L	6.22	1.6e-18	5.34
RSAD2	8.38	7.2e-37	8.59
CXCL10	9.79	4.9e-41	10.65
STAT1	2.32	2.7e-34	1.99

What this means for your project

• If your design compares groups of subjects, the analysis must aggregate to subject level. We do this by default and state it in the methods.
• If a previous analysis of your data used a per-cell test, it is worth re-running before submission — that is exactly what our re-analysis service is for.
• The full reasoning is written up in the pseudobulk guide.

→ Get a quote · single-cell sample · spatial sample · bulk RNA-seq sample